Build From Source

Build the Ollama MLX preview locally on Apple Silicon macOS when you want to inspect the source, avoid a binary download, or test a preview tag directly.

Before you start

These instructions build Ollama from the public preview source. They are intended for Apple Silicon Macs and MLX experimentation. For normal installation, use the latest signed preview DMG.

The app-bundle command below performs local ad hoc signing when no Apple Developer ID is configured. That produces a local test app, not a notarized redistributable DMG.

1. Install macOS build tools

Install Homebrew from brew.sh first if it is not already installed. Homebrew's installer also installs Apple's command line tools when they are missing.

Then install the build tools and download the Metal toolchain component:

brew install go cmake ninja node
xcodebuild -downloadComponent MetalToolchain

The important requirements are an Apple Silicon macOS environment, Go, CMake 3.24 or newer, Ninja, Git from Apple's command line tools, npm for the app UI, and Apple's Metal toolchain.

2. Clone the latest preview tag

Use the immutable preview tag for a reproducible build. The moving integration branch is useful for following current development, but a tag records the source state used for a published preview.

git clone \
  --branch preview/mlx-0.34.1-20260917 \
  https://github.com/pd95/ollama.git
cd ollama

Current source branch: integration v0.34.1 . Current preview tag: preview/mlx-0.34.1-20260917 .

3. Build and run the server

For a quick source test, build the local Ollama CLI/server binary with CMake, then start the server from the source checkout.

cmake -B build .
cmake --build build --parallel 8
./ollama serve

Leave the server running in that terminal. Use another terminal for ollama run, ollama create --experimental, or API tests.

4. Build the full app locally

To build dist/Ollama.app, use the Darwin build helper. On Apple Silicon, the arm64-only build is the practical local path.

./scripts/build_darwin.sh -a arm64 build app
open dist/Ollama.app

When APPLE_IDENTITY is not set, the helper ad hoc signs the app bundle and removes quarantine metadata so it can launch locally. It also writes dist/Ollama-darwin.zip and dist/ollama-darwin.tgz for local inspection.

Developer ID signing and notarized DMG packaging require Apple account credentials and are separate from this local app build. Do not redistribute the ad hoc signed app bundle.

5. Create local models

The preview source does not include model weights. Download model snapshots from their official Hugging Face repositories, accept the publisher license terms where required, and create local Ollama models.

Follow the model recipes for GPT-OSS, Apertus 1.0, Apertus 1.1 Mini, Apertus 1.5 Omni, and Gemma 4.
Apertus 1.5 users need to obtain the official swiss-ai/Apertus-v1.5-8B snapshot before importing it locally.